<p>Due to the poor feature extraction ability, weak anti-noise robustness, and generalization ability of traditional fault diagnosis methods, they cannot fully extract bearing fault features, which is prone to misdiagnosis and missed diagnosis. Aiming to address these problems, this paper proposes a novel feature fusion method combining an enhanced deep residual shrinkage network with global attention mechanism (DRSNGAM) and Swin Transformer (SwinT) for accurate fault diagnosis under complex working conditions. In the method, time–frequency analysis is adopted for data preprocessing and then to convert one-dimensional original vibration signals into two-dimensional time–frequency images. At the same time, the global key feature information and local detail feature information of the two-dimensional time–frequency images are extracted by the global feature extractor of the DRSNGAM and the local feature extractor of SwinT, respectively. Furthermore, the feature information is used to output the fault state identification results through convolution fusion, global average pooling, and fully connected layer. The effectiveness and superiority of the proposed method are validated on the two different public bearing datasets from Case Western Reserve University (CWRU) and Huazhong University of Science and Technology (HUST). The identification accuracy on the CWRU dataset is 99.89%, which is 0.97–3.59% points higher than that of the six advanced methods. Especially under complex working conditions with strong noise from – 4 to 8 dB, the identification accuracy of the proposed method is 88.39–99.49%. In addition, for the experiment conducted under different speed conditions on the HUST dataset, the identification accuracy is more than 98.66%.</p>